High Perf Browser

High Performance Browser Networking Framework

A systematic approach to web performance grounded in how browsers, protocols, and networks actually work. Apply these principles when building frontend applications, setting performance budgets, configuring servers, or diagnosing slow page loads.

Core Principle

Latency, not bandwidth, is the bottleneck. Most web performance problems stem from too many round trips, not too little throughput. A 5x bandwidth increase yields diminishing returns; a 5x latency reduction transforms the user experience.

The foundation: Every request passes through DNS resolution, TCP handshake, TLS negotiation, and HTTP exchange before a single byte of content arrives — each step adding round-trip latency. High-performance applications minimize round trips, parallelize requests, and eliminate unnecessary network hops. Understanding the protocol stack is the prerequisite for meaningful optimization.

Scoring

Goal: 10/10. Score by how many of the eight Quick Diagnostic rows pass, weighted toward the field metrics: 9-10 = all eight pass (the four field-metric rows in the green plus content-hashing, HTTP/2+, minimized render-blocking, and compression); 5-6 = the four field-metric rows pass but one or more transport/caching/compression rows fail; <=3 = any field-metric row is in the red. Always report the score, which diagnostic rows failed, and the specific fix for each.

The High Performance Browser Networking Framework

Six domains for building fast, resilient web applications:

1. Network Fundamentals

Core concept: Every HTTP request pays a latency tax — DNS lookup, TCP three-way handshake, TLS negotiation — before any application data flows. Reducing or eliminating these round trips is the single highest-leverage optimization.

Why it works: Light travels at a finite speed: a New York–London packet takes ~28ms one way regardless of bandwidth. These physics-level constraints cannot be solved with bigger pipes — only with fewer trips.

Key insights:

• TCP three-way handshake adds one full RTT before data transfer begins
• TCP slow start limits initial throughput to ~14KB (10 segments) in the first round trip — keep critical resources under this threshold
• Upgrade to TLS 1.3: it halves the handshake round trips of TLS 1.2 and enables 0-RTT resumption for returning visitors
• Head-of-line blocking in TCP means one lost packet stalls all streams on that connection
• Bandwidth-delay product caps in-flight data; high-latency links underutilize bandwidth

Code applications:

See references/network-fundamentals.md when tuning servers or diagnosing handshake latency — the full TLS 1.2-vs-1.3 RTT derivation, slow-start doubling table, initcwnd/BDP math, OCSP-stapling Nginx config, and the DNS cache hierarchy.

2. HTTP Protocol Evolution

Core concept: HTTP evolved from a simple request-response protocol into a multiplexed, binary system. Choosing the right protocol version and configuring it properly eliminates entire categories of performance problems.

Why it works: HTTP/1.1 forces workarounds (domain sharding, sprites, concatenation) because it cannot multiplex. HTTP/2 multiplexes but inherits TCP head-of-line blocking; HTTP/3 (QUIC over UDP) eliminates it. Each generation removes a bottleneck — and makes the previous generation's workarounds counterproductive.

Key insights:

• HTTP/1.1 allows one outstanding request per TCP connection; browsers open 6 per host as a workaround
• HTTP/2 multiplexes unlimited streams over one connection — domain sharding becomes counterproductive
• HPACK header compression in HTTP/2 cuts repetitive header overhead by 85-95%
• HTTP/3 (QUIC) eliminates TCP head-of-line blocking and enables 0-RTT resumption and connection migration
• Prefer 103 Early Hints over HTTP/2 Server Push (which over-pushes and is widely deprecated)
• Connection coalescing lets one HTTP/2 connection serve multiple hostnames sharing a certificate

Code applications:

See references/http-protocols.md when picking or migrating a protocol version — side-by-side HTTP/1.1-vs-2-vs-3 comparison, the step-by-step de-sharding migration, and why Server Push lost to 103 Early Hints.

3. Resource Loading and Critical Rendering Path

Core concept: The browser must build the DOM, CSSOM, and render tree before painting pixels: HTML → DOM → CSSOM → Render Tree → Layout → Paint → Composite. Any resource that blocks this pipeline delays first paint.

Why it works: CSS is render-blocking (no paint until CSSOM is ready) while JavaScript is parser-blocking (<script> halts DOM construction until it downloads and executes) — so each needs a different optimization strategy. Every blocking resource adds latency directly to time-to-first-paint.

Key insights:

• async downloads in parallel and executes immediately (use for independent scripts); defer downloads in parallel but executes after DOM parsing (use for most scripts)
• <link rel="preload"> fetches critical resources at high priority now; rel="prefetch" fetches likely next-navigation resources at low priority
• Inline above-the-fold CSS and async-load the rest to eliminate the render-blocking CSS request
• Fonts can block text rendering for up to 3s — use font-display: swap

Code applications:

See references/resource-loading.md when shaving first paint — the exact async/defer/module execution order, the full resource-hint decision tree, and the image/font (font-display, srcset, AVIF) playbook.

4. Caching Strategies

Core concept: The fastest network request is one that never happens. Layer caches — browser memory, disk, service worker, CDN, origin — to eliminate round trips for repeat visitors.

Why it works: Cache-Control headers tell the browser and intermediaries exactly how long a response stays valid; content-hashed URLs make aggressive immutable caching safe. Each cache hit eliminates a full network round trip.

Key insights:

• Cache-Control: no-cache still caches but revalidates every time; no-store never caches — don't confuse them
• ETag / Last-Modified enable conditional requests (304 Not Modified) that skip the body transfer
• Service workers provide a programmable cache layer that works offline (cache-first shell, network-first dynamic content)
• Misconfigured Vary headers cause CDN cache pollution — serve the wrong encoding or format to the wrong client

Code applications:

See references/caching-strategies.md when designing a cache policy — the full browser/SW/CDN/origin hierarchy, copy-paste service-worker cache-first vs network-first recipes, and the Vary pitfalls that pollute a CDN.